bilateral negotiation - tu delft
TRANSCRIPT
1
Vermelding onderdeel organisatie
January 29, 2009
1
Bilateral Negotiation
An overview of the theory of negotiation & negotiating software agents
Almen, HTI Winterschool
Koen Hindriks
Man-Machine Interaction Group, EEMCS
December 11, 2007 2
Outline
• A Classic Tale: Negotiating About an Orange
• Preferences and Utility Theory
• Bargaining Models and Solution Concepts
• Strategies and Concession Tactics
• Opponent Modeling & Learning
• Other Topics, or: What I Did Not Talk About
• The Future of Automated Negotiation
• Negotiation Session
December 11, 2007 3
A Classic Tale:Two Sisters Arguing Over an Orange*
* The story about the sisters' conflict over the orange has been attributed to Mary Parker Follett. See Deborah M. Kolb, 1995, The Love for Three Oranges, Or: What Did We Miss about Ms. Follett in the Library? 11 Negotiation J. 339. Also see: Roger Fisher & Danny Ertel, 1995, Getting Ready to Negotiate.
Two sisters arguing over a single available orange:
Key Concepts:• Conflict of interests• Distributional negotiation problem• Positional bargaining
I should have that orange. It's mine. No, I should have
it. It's mine.
I’m older, so I should have it. No, I should have it
because I’m younger.
December 11, 2007 4
A Classic Tale:Two Sisters Arguing Over an OrangeTwo sisters arguing over a single available orange:
Key Concepts:• Outcome space (Ω)• Status Quo• Utility 1
1
0 Utility youngest sister
Utility
oldest siste
r
No one gets
orange
Oldest sister gets orange
Youngest sister gets orange
December 11, 2007 5
A Classic Tale:Two Sisters Arguing Over an OrangeThe arguments are not decisive, but there is an alternative:
Key Concepts:• Making a concession, compromise• Expand outcome space• Fair outcome
This is going nowhere.
I agree.
Let’s compromise.
Let’s split the orange.
December 11, 2007 6
A Classic Tale:Two Sisters Arguing Over an OrangeThe arguments are not decisive, but there is an alternative:
Key Concepts:• Pareto optimal outcome• Pareto frontier 1
1
0 Utility youngest sister
Utility
oldest siste
r
Split the orange
2
December 11, 2007 7
A Classic Tale:Two Sisters Arguing Over an OrangeBut why do both sisters want the orange?
Key Concepts:• Underlying interests• Exchange of information
I want the peel for an orange marmalade I
am making for a school cooking project
I want to have the juice for a refreshing breakfast beverage.
You can have the juice. You can have
the peel.
December 11, 2007 8
A Classic Tale:Two Sisters Arguing Over an OrangeBut why do both sisters want the orange?
Key Concepts:• Inefficient outcome• Pareto dominated• Utopia, or Ideal point 1
1
0 Utility youngest sister
Utility
oldest siste
r One gets peel, other gets juice
December 11, 2007 9
Summary
• A negotiation is aimed at resolving a conflict of interests.
• Each negotiation is defined by its outcome space. Ideally, an agreement is reached that is Pareto optimal.
• During a negotiation, the outcome space is explored, and sometimes, may be expanded to find creative agreements based on the underlying interests.
Literature:
• H. Raiffa, 1982, The Art and Science of Negotiation, Belknap Press
• L.L. Thompson, 2005 (3rd ed.), The Mind and Heart of the Negotiator, Prentice Hall
Vermelding onderdeel organisatie
December 11, 2007
10
Preferences and Utility Theory
Von Neumann-Morgenstern Utility Theory
December 11, 2007 11
Where Do Utilities Come From?
• It is assumed that agent have preferences for outcomes, and thus are not completely indifferent about various outcomes.
• If preferences of an agent are both rational as well as well-behaved in a precise sense, then it can be shown that a utility function can be constructed that can be used for guiding the decision-making of that agent.
December 11, 2007 12
Predictable versusUnpredictable Preferences
Predictable preferences
• Most people prefer to have more rather than less money.
à Monotonic preference
à Logarithmic curve for money
Unpredictable preferences
• I like blue more than red, you like red more than blue.
à Varies from person to person
à Not quite as well-behaved
money
utility
red
utility
blue green …
• Higher > Lower Grade• Less > More Work
• Peugeot > Ford• Swim > Cycle Ride
3
December 11, 2007 13
Rational Preferences
Preferences are complete:
• It is assumed that an agent can state his preferences when confronted with two arbitrary potential outcomes.
• Formally, for all ω,ω’∈Ω: either ω ω’, or ω’ ω, or ω~ω’.
Preferences are transitive:
• If an outcome ω1 is preferred over
ω2, and ω2 is preferred over ω3,
then ω1 is preferred over ω3.
• Formally, for all ω1, ω2, ω3∈Ω:
ω1 ω2 and ω2 ω3 Ł ω1 ω3.
Preferences can be modeled by a (binary) preference relation . A preference relation is called rational if it is complete and transitive.
Intransitive Preferences& Money Pumps
A
C B1 Euro
1 Euro1 Euro
December 11, 2007 14
Modeling Risky Outcomes: Lotteries
• A lottery associates chances or probabilities with outcomes.
Notation: L = [p1, ω1 ; … ; pn, ωn].
Preferences are continuous:
• If an outcome ω’ is between ω and ω’’ in preference, then there is a probability p such that an agent is indifferent between the lottery [p, ω; 1-p, ω’’] and getting ω’ for sure.*
• Formally, if ω ω’ ω’’, then there is a p s.t. [p, ω; 1-p, ω’’] ~ ω’.
Example: Deal or No Deal.
• For a lottery involving monetary prizes, e.g.
[¼, 20; ¼, 100; ¼, 20.000; ¼, 40.000]
there is an amount of money, e.g. 5.000, such that:
[¼, 20; ¼, 100; ¼, 20.000; ¼, 40.000]~5.000.
Outcomes that involve risk can be modeled by lotteries. For each lottery, it is assumed there is a certainty-equivalent outcome.
* ω’ is called the certainty-equivalent outcome for lottery [p, ω; 1-p, ω’’].
December 11, 2007 15
Preferences are Monotone
Preferences are monotone:
• If an outcome ω is preferred over ω’, then an agent prefers a lottery with outcomes ω, ω’ that has a higher chance of getting ωover one with a lower chance of getting ω.
• Formally, [p, ω; (1-p), ω’] [q, ω; (1-q), ω’] whenever p≥q.
Preferences are assumed to be monotone. Intuitively, this assumption expresses that getting more of what you want is preferred over less.
Exceptions to monotonicity?Most people prefer life to death, but still some people like to exercise dangerous sports, such as BASE jumping, heli-skiing, diving, mountaineering, big wave surfing and bull riding. The chances of getting killed are significantly higher than staying at home listening to music, drinking a beer, or reading a good book.
December 11, 2007 16
Equally Valued Outcomes
Equally Valued Outcomes can be substituted:
• If an agent is indifferent between ω and ω’, then that agent is indifferent between a lottery that involves outcome ω over that same lottery with ω’ substituted for ω.
• Formally, [p, A; 1-p, C] ~ [p, B; 1-p, C] whenever A~B.
Equally valued outcomes may be substituted for each other.
The intuition here is that indifference with respect to two outcomes is independent from preferences regarding other outcomes.
December 11, 2007 17
No Fun in Gambling
Outcomes can be decomposed:
• Compound lotteries can be reduced using the laws of probability.
E.g., [p, A; [q, B; 1-q, C]] ~ [p, A; (1-p)q, B; (1-p)(1-q), C]
The axiom has also been paraphrased as stating that there can be
“no fun in gambling”. That is, an agent is indifferent between:
• Playing a lottery in which prizes are tickets for a 2nd lottery,
• Playing a single lottery that directly pays out cash
provided, off course, chances of winning cash are the same.
Lotteries that involve other lotteries can be reduced to simpler, non-composed lotteries.
December 11, 2007 18
Utility Theory
Theorem.
Suppose a preference relation satisfies all conditions listed above.
Then there exists a utility function U:Ωà ℜ such that:
U(ω)≥ U(ω’) iff ω ω
Example: Deal or No Deal:
• Suppose [¼, 20; ¼, 100; ¼, 20.000; ¼, 40.000]~5.000 and the preference relation satisfies all conditions above. Then the existence of a utility function U is guaranteed such that:
• ¼U(20)+¼U(100)+ ¼U(20.000)+ ¼U(40.000)= U(5.000).
The axioms of utility theory guarantee the existence of a utility function
U: ΩΩΩΩàààà ℜℜℜℜ that represents a rational, well-behaved preference relation.
4
December 11, 2007 19
Buying a SupercomputerIn practice, in complex negotiations, it is hard to construct a good utility function that matches the preferences of a party.
NCF/SARA negotiating about a HPC with IBM
Step 1: Identifying the main issues
Issues Value Space NCF
Weights
IBM
Weights
Planning
Delivery dates
6.2.1 System architecture
Number of TFlops
6.2.3 External and Internal Networking
Interconnect Network
6.2.4 Storage
Total storage
6.2.5 Software
Scheduler
6.2.7 Installation (Operatonal Aspects Section C)
Location
6.2.6 Maintenance and support - Related to SARA
Training
Startup: Onsite Support
Strategic
Use of Solution as a showcase for IBM
TCO (NCF) / Quote (IBM)
Quote Euro
Maintenance Euro
Total Costs Euro
Solution Utility
Overall NCF Utility
NCF IBM
Step 2: Identifying preferences & utility function
December 11, 2007 20
Summary
• A preference relation is a binary relation over an outcome space that models which outcomes are preferred over others.
• Preferences are rational if they are transitive and complete.
• A unique (Von Neumann-Morgenstern) utility function exists that models rational preferences that additionally satisfy continuity, monotonicity, substitutability, and decomposability.
Literature:
• R.D. Luce and H. Raiffa, 1957, Games and Decisions: Introduction and a Critical Survey, Dover Publications
• R.L. Keeney and H. Raiffa, 1976, Decisions with Multiple Objectives, Cambridge University Press
Vermelding onderdeel organisatie
December 11, 2007
21
Bargaining Models andSolution ConceptsAxiomatic Foundations for Negotiation
December 11, 2007 22
Abstract Model of Negotiation
• A bargaining problem is a pair ⟨Ω, d⟩ with Ω an outcome space, and d the disagreement point.
• A solution for bargaining problems of the form ⟨Ω, d⟩ is an outcome ω=F(Ω,d)∈Ω.
• F(Ω,d) can be interpreted as a prediction, or recommendation, for the problem ⟨Ω, d⟩.
Two parties that negotiate need to make a joint decision to settle on a final outcome. Can we predict the outcome that such agents will agree on?
Utility
of agent 2
Dictatorialsolution
Utopia, orIdeal point
Utility of agent 1
ParetoOptimalFrontier
Outcome space= feasible region,
bounded byaxes and
Pareto frontier
Disagreement point,or conflict outcome
December 11, 2007 23
Which Solution Should We Select?
• Characterize desirable properties of solution.
• Ideally, but not necessarily, this would give a unique solution point for each bargaining problem.
• Axiomatic approaches to define properties.
• We discuss the axioms proposed by Nash (1950).
December 11, 2007 24
Axiom 1:All Gains Should Be Exhausted
• Pareto-Optimality:
A solution ω=F(Ω,d) should be
Pareto-optimal, i.e. such that
there is no ω’∈Ω and ω’> ω.
Most solution concepts proposed satisfy this property.*
* The Egalitarian solution satisfies weak Pareto optimality, i.e. there is no outcome that is preferred by all negotiators to the Egalitarian solution. A natural extension of this solution concept that does satisfy Pareto optimality exists.
5
December 11, 2007 25
Axiom 2:Fully Symmetric Bargaining Problems
• Symmetry:
If negotiators have completely
symmetric roles, the solution
should yield them equal utility,
i.e. the solution F(Ω,d)=⟨x1, x2⟩
should satisfy x1=x2.
Does not allow for external factors that are not modeled.
December 11, 2007 26
Axiom 3Scale Invariance
• Scale Invariance:
The solution should be invariant
under “scale transformations”, i.e.
for all positive linear transformations
λ we have λ(F(Ω,d))=F(λ(Ω), λ(d))
Does not allow for interpersonal comparison of utilities.
December 11, 2007 27
Axiom 4Independence of Irrelevant Alternatives
• Contraction Independence:
If Ω’⊆Ω and F(Ω,d)∈Ω’,
then F(Ω’,d)=F(Ω,d).
Does not allow to take into account that one negotiator has
more or less options available to choose from in his favor.
December 11, 2007 28
The Nash Solutionto a Bargaining Problem
• It can be shown that there is a unique outcome that satisfies all axioms proposed by Nash, called the Nash solution.
• The Nash solution is that outcome ω=⟨x1, x2⟩ ∈Ωthat maximizes x1⋅x2.
December 11, 2007 29
Many Alternative Solutions Proposed
Best Known Solutions:
• Nash Solution:
outcome ω=⟨x1, x2⟩ that maximizes x1⋅x2
• Kalai-Smorodinsky Solution:
Pareto-optimal outcome ω on line that connects disagreement and ideal point
• Egalitarian Solution:
Weak Pareto-optimal outcome ω=⟨x1, x2⟩such that x1=x2
Many Other Solutions:
•Dictatorial Solution
•Raiffa Solution
•Kalai-Rosenthal Solution
•Lexicographic Egalitarian
•Perles-Maschler Solution
•The Equal Area Solution
•The Equal Loss Solution
•The Utilitarian Solution
•The Yu Solution
December 11, 2007 30
Summary
• A bargaining problem consists of an outcome space and a disagreement point
• A solution to a bargaining problem recommends a unique outcome as the final agreement
• Nash proposed to axiomatically characterize solution concepts
• Many solution concepts proposed in literature
Literature:
• R.D. Luce and H. Raiffa, 1957, Games and Decisions: Introduction and a Critical Survey, Dover Publications
• W. Thomson, 1994, Cooperative Models of Bargaining, in: Handbook of Game Theory, Volume 2, Elsevier
6
Vermelding onderdeel organisatie
December 11, 2007
31
Strategies and Concession Tactics
Alternating Offers Protocol with Incomplete Information
December 11, 2007 32
Joint Exploration of Outcome Space
• Axiomatic approaches typically characterize possible outcomes of a negotiation, assuming complete information.
• In closed negotiations, negotiators at best have partial information about their opponent’s preferences.
• For a good reason: Revealing information to an opponent may result in exploitation.
• Reaching an agreement requires negotiators with incomplete information to jointly explore the outcome space by exchanging offers, i.e. outcomes proposed to the other party.
The Nash solution, and, more importantly, the Pareto frontier, cannot be computed when both parties only have incomplete preference information.
December 11, 2007 33
Fundamental Problem of Negotiation
What is the optimal strategy to achieve the following:
• Maximize own outcome.
Requires outcome with high own utility, i.e. that is individually rational. Rule of Thumb: Start negotiation and check whether offer that is best for self is acceptable.
• Maximize chance of reaching an agreement.
Requires outcome with acceptable utility for opponent, i.e. resolving the conflict of interest.
A dilemma faced by any negotiator is how to simultaneously achieve two typically conflicting goals: maximize own outcome and chance of a deal.
December 11, 2007 34
Negotiation Decision functions
Different families of
negotiation heuristics:
• Time-dependent tactics
• Behaviour-dependent tactics
The complexity of negotiation with incomplete information does not allow for establishing game-theoretic optimal equilibrium strategy results. Instead negotiation heuristics are proposed in the literature.
?
A tactic is used to determine a next offer.
December 11, 2007 35
Time Dependent Tactics
• A time dependent tactic determines the next offer to be proposed as a function αa(t) of time:
Various time-dependent functions are used:
• A polynomial function:
• An exponential function:
Time dependent tactics take deadlines and discounting of future rewards into account.
−−+
−+=→
increasingVifminmaxtmin
decreasingVifminmaxtminx
aaaaa
aaaaa
t
ba)())(1(
)()(
α
α
β
α
1
max
max ),min()1()(
−+=
a
a
aaa
t
ttkkt
akat
att
aet
ln)
max
)max,min(1(
)(
β
α
−
=
December 11, 2007 36
Time Dependent Functions
0 10 2 0 30 40 50
0
2 0
4 0
6 0
8 0
10 0
tb
max
kb(=5% )
maxb
m inb
B ou lw are (be ta=0 .025 )
B ou lw are (beta=0 .1 )
Linea r(be ta=1 )
C onceder(be ta=10)C onceder(be ta=50)
bu
ye
r o
ffe
r
tim e
Decision Parameters:
•Reservation values: mins, maxb
•Initial offer level: k
•Deadline (time available): tmax
•Rate of concession: β/µ/γ/δ,…
7
December 11, 2007 37
Behavior Dependent Tactics
• A behavior dependent tactic determines the next offer to be proposed as a function ψ(tn) of the last proposed offer tn:
Various functions ψ(tn) are used:
• Average tit-for-tat:
• Relative tit-for-tat:
Behavior dependent tactics take the behavior of the opponent into account measured relative to the agent’s own utility space, giving rise to tit-for-tat strategies.
)),),(min(max(1 aa
n
t
ba maxmintx n ψ=+→
1
2
)( −
−
→→
→= n
n
n
t
bat
ab
t
ab
n xx
xt
γ
ψ
1
22
2
)( −
+−
−
→→
→= n
n
n
t
bat
ab
t
ab
n xx
xt
δ
δ
ψ
December 11, 2007 38
Monotone Concession Protocol
• Rule: Players are not allowed to make offers which have a lower utility for their opponent than their last offer.
• If the minimum concession is required to be >0, then a negotiation is guaranteed to terminate.
• Agents accept offer on table if utility exceeds that of own last offer or that of a viable counter-offer.
• Example: Zeuthian Strategy (cf. literature).
Note:
• In multi-issue negotiations with incomplete information about opponent preferences, it is not always possible to make monotonic concessions (due to “compatible issues”)
• Neither is it best to monotonically make concessions, as we will see later.
A monotone concession protocol requires agent’s to make concessions at each step, i.e. increase their opponent’s utility at each negotiation move.
December 11, 2007 39
Smart Meta-Strategy
• Key idea: To better meetthe demands of an opponentit is not always required toconcess utility oneself.
• Propose alternative offer that ismore similar to that of opponent’slast offer with same own utility.• Only concess if not possible.
• Result: Issue trade-offs. Example: Propose ⟨Medium Quality ↓, Low Price ↑, 10 days =⟩ instead of ⟨High Quality, Medium Price, 10 days⟩
• See: Trade-Off Strategy (cf. literature)
Using domain knowledge it may be possible to improve the efficiency of a strategy to the benefit of the agent itself as well as of its opponent.
?
Iso-utility curve
December 11, 2007 40
Evaluating Strategies
Process-Oriented:
• Cost-Effective: How many steps are required to reach agreement?
• Robustness: Is it possible to exploit the strategy?
• Negotiation move types: Is strategy sensitive to opponent?
Outcome-Oriented:
• Successful: Is an agreement reached?
• Efficiency: Is agreement Pareto optimal?
• Fairness: Is agreement close to Nash? Kalai-Smorodinsky?
A number of criteria are available to judge the quality of a strategy.
December 11, 2007 41
Evaluation:Dynamics of Negotiation Process
Trade-Off versus ABMP strategy on the City-vs-AMPO domain (Raiffa)
Unfortunate step
Inefficient
December 11, 2007 42
Evaluation:Dynamics of Negotiation Process
Trade-Off versus Random Walker strategy on the City-vs-AMPO domain
Sensitive to opponent
Nice moves
8
December 11, 2007 43
Summary
• Due to incomplete information about opponent preferences, it is not possible to define “optimal” strategies in any precise sense.
• Various negotiation heuristics have been proposed, including time and behavior dependent tactics.
• Efficiency can be improved by using domain knowledge and proposing an offer on a utility iso-curve more similar to the opponent’s last offer.
Literature:
• P. Faratin, C. Sierra, N.R. Jennings, 1997, Negotiation Decision Functions for Autonomous Agents, in: Int. Journal of Robotics and Autonomous Systems, 24(3-4), 159-182
• P. Faratin, C. Sierra, N.R. Jennings, 2002, Using Similarity Criteria to Make Issue Trade-Offs in Automated Negotiations, in: Artificial Intelligence, 142, 205-237
• K.V. Hindriks, C.M. Jonker, D. Tykhonov, 2007, Analysis of Negotiation Dynamics, in: Cooperative Information Agents XI, LNCS 4676, 27-35
• J.S. Rosenschein, G. Zlotkin, 1994, Rules of Encounter, MIT Press
Vermelding onderdeel organisatie
December 11, 2007
44
Opponent Modeling and Learning
Improving negotiated agreements by estimating opponent preferences.
December 11, 2007 45
Learning Opponent Preferences
• The strategies and tactics discussed so far are blind for opponent preferences.
• Various techniques have been proposed to learn features and construct an opponent model.
• Can information about opponent preferences be learned in single-instance negotiation?
• Single-instance negotiation typical for e-commerce.
• Hard because only bids exchanged can be used, no previous history can be assumed.
Various features of an opponent’s strategy can be learned, e.g. type of concession tactic, reservation value. We discuss learning opponent preferences.
December 11, 2007 46
Learning: Structural Assumptions
• Linear additive utility functions:
• Sufficient to learn ranking of weights:
• Shape of Evaluation functions (utility of issues):
Downhill,
Uphill, and
Triangular
∑=
∈=n
i
tiiit bxewbu1
)()(
( )12
+=
nn
rw
j
i
i
0
1
ei(xi)
xi
hei,1 h
ei,2 h
ei,3
hei,k-2 h
ei,k-1 h
ei,k
In order to be able to learn opponent preferences we have to exploit certain structural features of negotiation problems.
December 11, 2007 47
Learning: Rationality Assumptions
• Opponent starts by
proposing its best bid.
• Opponent uses a
concession-based tactic
• Probability distributions are associated with possible tactics.
In order to be able to learn opponent preferences we have to assume agents are (at least to some extent) rational.
0
P(b0|hj)
P(b1|hj)
P(b2|hj)
P(bt|hj)
u’(b1)u’(b2) u’(b0)
u(bt|hj)
December 11, 2007 48
Bayesian Learning
• Using Bayes’ Rule the probabilities associated with hypotheses regarding weights and evaluation functions can be updated for every new bid bt:
• An estimate of the utility an opponent associates with a bid can be computed using the computed probabilities:
Each (new) bid provides new information that is used to update the model of the opponent’s preferences using Bayes’ rule.
( ) ( ) ( )( ) ( )∑
=
=m
k
ktk
jtj
tj
hbPhP
hbPhPbhP
1
( ) ( ) ( )∑ ∑= =
∈=H
j
n
i
tiiijt bxewhPbu1 1
9
December 11, 2007 49
Using an Opponent Model
An opponent model:
• enables choosing an offer that has utility x for the negotiator and maximizes the utility of the opponent (which helps to increase the chance of acceptance)
• enables to select an offer that increases the utility of both parties if previous offers have deviated from the Pareto frontier, which is typical. I.e. it allows making a fortunate move.
An opponent model can be used to choose a next offer in bilateral negotiation.
December 11, 2007 50
Defining an Opponent-based Strategy
Strategy using opponent model:• Use a tit-for-tat tactic, i.e.
match negotiation move that opponent just made.
• Propose Pareto-efficient offers, i.e. go “straight up” (agent A) to the Pareto frontier in order to
maximize own utility.
• Many possible variants.
December 11, 2007 51
Summary
• It is possible to learn opponent preferences even in single-instance negotiations by making some rationality and structural assumptions.
• Efficiency can be improved by using learnt opponent preferences and by proposing an offer that increases utility of both parties.
Literature:
• D. Zeng, K. Sycara, 1998, Bayesian Learning in Negotiation, in: International Journal of Human–Computer Studies 48, 125–141
• R. Lin, S. Kraus, J. Wilkenfeld, J. Barry, 2006, An Automated Agent for Bilateral Negotiation with Bounded Rational Agents with Incomplete Information, ECAI 2006, 270-274
• K.V. Hindriks, D. Tykhonov, 2007, Opponent Modelling in Automated Multi-Issue Negotiation Using Bayesian Learning, Submitted.
Vermelding onderdeel organisatie
December 11, 2007
52
Other Topics, or:What I Did Not Talk AboutAutomated negotiation is a very active and broad area of research.
December 11, 2007 53
OtherTopics
Auctions
December 11, 2007 54
OtherTopics
10
December 11, 2007 55
Summary
• Topics discussed include: Decentralized, Bilateral (one-to-one) & closed versus open negotiation. Multi-issue, additive utility functions. Alternating offers protocols without time pressure. Automated negotiating software agents. Human-machine negotiation.
• Topics not discussed include: Multi-lateral negotiation. Auctions. Human negotiation, emotion, personality & culture. Argumentation-based and qualitative models of negotiation. Deadlines. Complex, non-linear utility functions.
Literature:
• V. Krishna, 2002, Auction Theory, Academic Press
• I. Rahwan, S.D. Ramchurn, N.R. Jennings, P. McBurney, S. Parsons, L. Sonenberg, 2003, Argumentation-based Negotiation, in: Knowledge Engineering Review, 18(4), 343-375.
• L.L. Thompson, 2005 (3rd ed.), The Mind and Heart of the Negotiator, Prentice Hall
Vermelding onderdeel organisatie
December 11, 2007
56
The Future of Automated NegotiationAutomated negotiation has potential applications in the near future in electronic commerce and marketplaces, and negotiation support systems.
December 11, 2007 57
The Future of Automated Negotiation
Strategies:• Optimal strategies for negotiation with incomplete information.
Learning:• negotiation tactics, opponent models, deadlines, …
Qualitative negotiation models:• Qualitative preference models, extensions of the alternating offers protocol that allow for additional forms of information exchange.
Preference modeling:• Elicitation techniques, utility spaces with issue dependencies.
Research Challenges
December 11, 2007 58
The Future of Automated Negotiation
Negotiation support systems:• Human-machine interaction, • Integration of emotion & personality, negotiation styles, culture• Analyzing domains and real-world cases
Electronic Marketplaces:• Online negotiation, shopping bots (pre-negotiation)• Recommender systems, personalized negotiation assistance,
information selection & provision.• Empirical analysis.
Applications